Designing the Freshness Gate

Your ask-the-web agent handles both "what's the atomic number of carbon" and "who won last night's game." Both queries return relevant, high-confidence sources from the reranker.

1. Explain why confidence and freshness are two different questions, with a concrete case where a query could pass one gate and fail the other.
2. List three concrete signals you'd use to decide whether to force a live search versus trust a cached result, and say which is cheapest to compute and which is most reliable.
3. Your freshness gate forces a live search on the "atomic number of carbon" query, every time, because the query classifier is miscalibrated and routes it to the news path. What's the cost of this specific failure mode, and how would you catch it before a user does?